humanoids gain human-like agility with ASAP framework

Humanoid robots look like Olympic gymnasts in simulations, but the moment they hit real-world concrete, they often turn into stiff, clumsy toddlers. Bridging this gap between virtual physics and messy reality has been one of robotics' biggest headaches.

A research team has introduced ASAP, short for Aligning Simulation and Real Physics, a two-stage framework designed to teach humanoids agile, whole-body dynamic skills without losing their speed.

Ditching conservative training wheels

Historically, engineers used domain randomization to prepare robots for the real world. By throwing random frictions and weights at the simulated robot, they made it robust, but at a painful cost: the robot became overly cautious and lost its agility.

ASAP takes a different route. First, it pre-trains a whole-body motion tracking policy inside a simulator using human motion capture data.

Next, researchers deploy that policy on hardware, collect real-world performance data, and train a delta action model. Think of this model as an offset calculator that measures exactly how real-world motors and friction deviate from the virtual world.

Finally, ASAP plugs this delta model back into the simulator to fine-tune the original policy against real-world physics.

Tested on the Unitree G1

The team tested the method across multiple virtual environments, including IsaacGym to IsaacSim and IsaacGym to Genesis. More importantly, they ran it on the Unitree G1 humanoid robot in the real world.

Compared to classic system identification, domain randomization, and delta dynamics baselines, ASAP sharply reduced tracking errors while keeping motions fluid and dynamic.

The real Unitree G1 managed to execute high-agility maneuvers that typically fail or look broken when transferred directly from simulation.

Why it matters

If you build humanoid software, ASAP changes the math on sim-to-real training. You no longer have to pick between a robot that breaks on contact or one that moves like a slow-motion statue. It offers a practical recipe to get expressive, agile movement onto real hardware quickly.

For the robotics industry, dynamic whole-body coordination is the real barrier between scripted lab demos and commercial utility. Humanoids in logistics or construction cannot afford to pause and rebalance on every uneven tile.

We think the clever part here is correcting the actions rather than trying to build a theoretically perfect simulation engine. That said, real robots suffer from motor overheating and mechanical wear over time, so how well these residual action models adapt across long operating shifts remains an open question.